CSCE 121 Culture Report 1

Efficient Visual Object Recognition

Robotics, artificial intelligence, and autonomy are rapidly evolving fields in modern technology. New hardware is being developed to send visual imagery to the software operating system to interpret and make decision. For example, the DARPA Urban Challenge race encourages engineers to develop vehicles capable of navigating streets by themselves. The accepted practice of object recognition has been to match a given image with a similar one that has been stored in a database. Such algorithms rely on huge databases for innumerable objects, and are therefore heavy on computer resources and difficult to accomplish in real-time, real-world situations.

According to IEEE Computer Magazine , Engineers at MIT's Computer Science and Artificial Intelligence Laboratory have been working on a new form of object recognition that identifies basic shapes and components of a larger, whole object—similar to how the human brain identifies objects: an unknown object's identity could be determined by the characteristics of its components, allowing the system to learn from what it has seen and more accurately identify similar objects in the future. For example, such a system would be able to identify a vehicle by identifying its four wheels, a chassis, and motion in a particular direction. This system is even able to recognize components (and therefore entire objects) with busy backgrounds. An added benefit to this method of object-recognition is its efficient use of resources. Therefore, MIT's new algorithms would allow faster recognition of known objects and accurately deduced recognition of unfamiliar objects.

I believe that object recognition systems have many practical applications, the most obvious of which is self-driven vehicles. Faster, more reliable object recognition algorithms would allow the vehicle's system to identify and avoid obstacles more readily. This advantage would make autonomous vehicles safer and more dependable. Such vehicles would be most beneficial to unmanned military operations: vehicles could be better trusted to take care of themselves while keeping our troops out of hamr's way. This object recognition system could have useful applications in engineering. Such a system would help engineers, mechanics, and technicians find problems/flaws by comparing a pre-defined virtual model of a building, bridge, engine, etc. with the actual recognized object. If a component is slightly out of alignment with what the computer knows to be correct, it could bring the problem to the designer's attention and perhaps provide feedback on how to remedy the problem.

The system could also be used in a somewhat reversed order: 3-Dimensional models require binocular vision: the perceived "jump" of an object between two cameras placed at a known distance apart from each other can give the precise distance to that object. This new component-based object recognition system might be able to generate a 3-dimensional model by identifying components of known sizes and determining the relative distances between them. Similar methods are probably already being used to construct models of objects using only a single image and a few known measurements, but they probably take a long time to construct. An automated 3D model generator would aid in the development of more realistic virtual environments in video games and enhanced street view in Google Maps.

This article raises one major question in my mind: Why has no one attempted this type of object recognition before? If artificial intelligence is supposed to be modeled after what we know about the human brain and our own intelligence, software engineers should have been able to create bottom-up object recognition algorithms much sooner. Perhaps either the digital imaging technology or computer language abstraction has only recently advanced to the point where we can take advantage of such advanced object processing. Simultaneously identifying several components of an object (or multiple objects) in an image must require a lot of careful programming and processing.